[Beowulf] WRF model on linux cluster: Mpi problem

[Peter Kjellström]

I agree with Joachim, somethings wrong. A decent network shoult give you < 35 
us latency and > 70 MiB/s bandwidh at larger packet sizes. Here are some 
things to check:

* if you use e1000, set interruptthrottlerate=0

* if you for speed and duplex make sure it's a 100% forced configuration, 
don't mix in autoneg in any way

* have a look at ifconfig (error numbers) and check dmesg for ugliness

* test between some other nodes, is it the same?

good luck,
 Peter

On Wednesday 06 July 2005 10.05, Joachim Worringen wrote:
> Federico Ceccarelli wrote:
> > Hi,
> >
> > I did the Pallas benchmark...after removing openmosix...here are the
> > ping-pong and ping-ping results...for 2 processes
> > What do you think about them?
> > Why the bandwidth is raising and decreasing many times as the #bytes
> > grow?
>
> The latency is quite high and grows too fast from 0 to 8 bytes. But most of
> all, the bandwidth should be constant with at least 80MB/s, better >100MB/s
> for large messages. Something is wrong, probably with your ethernet setup.
> I remember such effects for half/full-duplex mismatch with 100Mb auto
> negotiation problems between NIC and switch.
>
>   Joachim

-- 
------------------------------------------------------------
  Peter Kjellström               |
  National Supercomputer Centre  |
  Sweden                         | http://www.nsc.liu.se
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: not available
Url : http://www.scyld.com/pipermail/beowulf/attachments/20050711/4b0a6b95/attachment.bin